Lightning Talk: From Pretrained To Personal: Privacy-First... Daniel Holanda Noronha & Iswarya Alex

This talk explores the advancements in using PyTorch on AI PCs for meaningful model fine-tuning while maintaining data privacy. The speakers, Daniel Holanda Noronha and Iswarya Alex from AMD, discuss the design of on-device fine-tuning pipelines tailored for enterprise scenarios, particularly in regulated sectors like healthcare, government, and finance. They cover essential topics such as the selection of efficient pre-trained models and the implementation of PyTorch optimizations to facilitate personalization on large private datasets. The session also highlights practical fine-tuning techniques including supervised fine-tuning, LoRA, and QLoRA, demonstrating how mixed-precision training can enhance efficiency without compromising security. Ultimately, the talk presents a cloud-free, privacy-first fine-tuning approach that transforms AI PCs into secure personalization engines for enterprise applications.